Multitask radioclinical decision stratification in non-metastatic colon cancer: integrating MMR status, pT staging, and high-risk pathological factors.

Introduction
Colorectal cancer (CRC) is the third most common cancer worldwide and ranks second in cancer-related mortality [1]. Due to anatomical position and risk factors, the incidence of colon cancer is significantly higher than that of rectal cancer, accounting for approximately two-thirds of all CRC cases [2]. Based on the M-stage classification of the American Joint Committee on Cancer (AJCC) staging system, colon cancer can be categorized into non-metastatic (stages I–III) and metastatic (stage IV) disease [3]. Currently, the primary treatment modality for non-metastatic colon cancer (NMCC) remains curative surgery. However, postoperative prognosis varies significantly, with 20%−30% of patients experiencing tumor recurrence or metachronous distant metastasis despite standard surgical treatment [4].
Studies found that patients with deficient mismatch repair (dMMR) or microsatellite instability-high (MSI-H) CRC do not benefit from adjuvant chemotherapy [5, 6]. However, the NICHE trial have demonstrated that dMMR tumors show high pathological response rates to neoadjuvant immunotherapy, with pathologic complete response (pCR) rates of 100% and 95%, respectively [7, 8]. In 2017, the U.S. Food and Drug Administration (FDA) approved Pembrolizumab for the treatment of dMMR solid tumors, significantly improving survival rates for patients with dMMR colon cancer [9]. Meanwhile, neoadjuvant chemotherapy (NAC) is increasingly used in patients with proficient Mismatch Repair (pMMR) tumors, particularly in stage T4 colon cancer. FOxTROT trials have shown that NAC can improve local tumor control and potentially reduce the risk of metachronous distant metastasis [4, 10]. Furthermore, according to the 2024 National Comprehensive Cancer Network (NCCN) guidelines for colon cancer, patients with pMMR and pathological stages T1-3 must be further evaluated for high-risk features (e.g., histological differentiation, lymphovascular invasion, perineural invasion).The patient’s risk of recurrence and metastasis, along with the effectiveness of their prognosis, is intimately tied to these factors, which also decide whether additional chemotherapy is required [11]. Despite those treatment approaches for colon cancer become increasingly standardized, the effectiveness of these therapies varies by individual patient, making preoperative stratification crucial for selecting the most appropriate treatment strategy.
Currently, the gold standard for NMCC stratification is pathological evaluation and molecular marker testing, which are invasive, difficult to repeat and the accuracy is affected by sampling errors [12, 13]. Contrast-enhanced computed tomography (CECT) is a non-invasive, cost-effective imaging modality for preoperative evaluation in colon cancer. Many studies have shown that CECT features are closely correlated with tumor aggressiveness, molecular biomarkers, and prognosis in colon cancer [14–16]. Additionally, with the advancement of computer science, machine learning and deep learning are increasingly employed to predict tumor characteristics and patient outcomes. However, machine learning based on radiomic features often requires labor-intensive lesion segmentation and feature processing, while deep learning demands large sample sizes for training and lacks interpretability [17, 18]. Moreover, most existing studies focus on single aspects of tumor evaluation, and the potential for using CECT to guide global decision stratification for NMCC patients has not been fully explored or realized. Therefore, there remains an unmet need for simplified, interpretable, and clinically actionable models for NMCC stratification.
In summary, this study aims to develop a multitask global decision support system by identifying CECT-based features and integrating clinical parameters to enable precise stratification of NMCC patients. The proposed system seeks to provide preoperative, individualized treatment recommendations for NMCC, optimizing therapeutic strategies and ultimately improving patient survival and quality of life.

Methods

Study patients
This retrospective study was approved by the hospital’s Institutional Review Board (IRB) and waived the requirement for written informed consent (IRB number: KY2024052). Patients diagnosed with colon cancer and treated surgically between January 2021 and December 2023 at Institution I, and between 2013 and 2016 at Institution II, were included. Although there is a temporal difference in patient enrollment periods between the two centers, both institutions adhered to the same NCCN guidelines for colon cancer during their respective study periods, with key pathological assessment criteria and surgical techniques remaining consistent throughout the study period. The inclusion criteria were as follows: (a) postoperative pathology confirmed colon cancer; (b) abdominal CECT scan within two weeks before surgery; and (c) no prior treatment before surgery. The exclusion criteria were: (a) incomplete clinical or pathological data; (b) incomplete or poor-quality CECT images; (c) inability to identify the primary lesion on CECT images; (d) presence of synchronous distant metastasis confirmed by imaging or other diagnostic methods prior to surgery; and (e) history of colon cancer or other malignancies (Fig. 1).

Patient stratification
All pathological results were retrospectively reviewed from the patients’ postoperative tumor pathology reports. Patients were stratified stepwise through three task layers—MMR status, pathological T (pT) staging, and the presence of risk factors (e.g., histological differentiation, lymphovascular invasion, perineural invasion, bowel obstruction or perforation)—based on their treatment modalities (Fig. 2a). Detailed stratification information and pathological assessment information are included in Supplementary Material II.

CECT acquisition
All patients underwent standard CECT of the entire abdomen within two weeks prior to surgery. Detailed equipment information and scanning parameters are provided in Supplementary Material Ⅱ and Table S1. Patients received intravenous non-ionic iodinated contrast agents, including Iopromide 370, Iohexol 300, Iopamidol 370, Iomeprol 300, or Ioversol 350, administered at a rate of 2.5–3.5 mL/s, with a total dose of 1.5 mL/kg body weight. Arterial phase scans were triggered 6 s after the CT attenuation of the region of interest (ROI) placed over the aorta reached the threshold enhancement value, followed by portal venous phase scans 25–30 s later. All CT image data were transferred to the PACS system in DICOM format for evaluation by radiologists.

Clinical characteristics and CECT signs assessment
Clinical characteristics were collected from medical records, including seven key variables: gender, age, height, weight, body mass index (BMI), pre-treatment carbohydrate antigen 19 − 9 (CA19-9), and carcinoembryonic antigen (CEA). CECT signs contain primary tumor signs and suspicious lymph node (SLN) signs (Fig. 3). The detailed information of CECT signs can be found in Supplementary Material II.

Feature selection
A subset of 50 cases was selected for consistency evaluation of categorical and continuous data using Cohen’s Kappa and intraclass correlation coefficient (ICC), respectively, to assess intra- and inter-rater agreement for all CECT signs. CECT signs with Kappa or ICC values greater than or equal to 0.75, indicating good agreement, were included in the study. Both the retained CECT signs and clinical features were subsequently processed. For some continuous variables with minor missing data, mean imputation was applied. To eliminate the influence of scale, all feature values were standardized using the Z-Score method (Z = (X − mean)/standard deviation) before feature selection. Subsequently, five feature selection methods—Least Absolute Shrinkage and Selection Operator (Lasso), Boruta, Max-Relevance and Min-Redundancy (mRMR), Relief, and eXtreme Gradient Boosting - Recursive Feature Elimination (XGBoost-RFE)—were employed in each of the three task layers to rank feature importance (Fig. 2b). Following the empirical rule of 10 events per variable (EPV) [19], the number of features retained was determined by 10% of the sample size in the smallest treatment group. Features retained from all three task layers were then combined to form a total feature set. The detailed information on feature selection algorithms can be found in Supplementary Material Ⅱ.

Machine learning model construction
The machine learning modeling process was conducted using a commercial interactive multimodal research platform (Deepwise Healthcare, China, http://keyan.deepwise.com/), which integrates the Python (version 3.0.1) and scikit-learn (version 0.22) packages for medical data analysis. We first randomly split the dataset for each task layer into a training-validation set and an independent test set in a 7:3 ratio. For the training-validation set, we employed a 5-fold cross-validation strategy to train and validate the models, with the average results from the 5 folds used as the final outcome. To ensure comparability across different models within the same task layer, Logistic Regression was used as the baseline model, and a range of 100 random seed values (1–100) were tested to select the optimal seed, which was then fixed for subsequent analyses. With the total feature set, models were built for each of the three task layers using 14 machine learning algorithms: Logistic Regression, Decision Tree, Random Forest, K-Nearest Neighbors, Support Vector Machine (SVM), Linear Support Vector Classification (Linear SVC), Linear Discriminant Analysis (LDA), Stochastic Gradient Descent (SGD), Multilayer Perceptron, Gradient Boosting, Adaptive Boosting (AdaBoost), eXtreme Gradient Boosting (XGBoost), Bernoulli Naive Bayes (Bernoulli NB), and Gaussian Naive Bayes (Gaussian NB) (Fig. 2c). During model training, a combination of grid search and random search was used to automatically fine-tune hyperparameters for different model types, and the process was repeated 20 times to ensure model stability. Furthermore, we also employed the same methodology to directly construct a multi-class model based on the total feature set from three tasks, utilizing the same 14 machine learning algorithms for comparative analysis. Details on the machine learning algorithms are provided in Supplementary Material Ⅱ.

Statistical analyses
All statistical analyses were performed using Python (version 3.12.5). The Shapiro-Wilk test was used to assess the normality of the distribution for continuous variables. Continuous variables were presented as mean ± standard deviation for normally distributed data, or as median and interquartile range for non-normally distributed data. Categorical variables were expressed as frequencies and percentages. Differences between groups were compared using the Student’s t-test or Mann-Whitney U test for continuous variables, and the chi-square test or Fisher’s exact test for categorical variables. The area under the receiver operating characteristic curve (AUC-ROC) and its 95% confidence interval (CI) were calculated as the primary metrics to evaluate the performance of the models. Model accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1-score were calculated based on the threshold of 0.5. Due to the imbalance in sample sizes across different categories, a weighted Matthews correlation coefficient (MCC) was utilized to evaluate the overall performance of the multitask hierarchical model and the multi-class model in global decision stratification. Statistical significance was set at a two-tailed p-value of < 0.05. All analyses were performed independently for each of the three task layers (MMR status, pT stage, and risk factors). The statistical results were used to validate and complement the findings of the machine learning models.

Results

Patient characteristics
Patients from two institutions were pooled, resulting in a total of 372 NMCC patients included in the study. Institution I contributed 293 patients (median age, 61.19 years; 152 men), and Institution II contributed 79 patients (median age, 62.67 years; 46 men). Detailed information of patient characteristics is provided in Supplementary Material II and Tables S7-9.

Feature selection
After consistency analysis, 32 features were evaluated, and the arterial phase density values of the primary tumor and the standard deviation of venous phase density were excluded. Although the intra- and inter-reader Kappa values for the feature “Period of peak density” were both below 0.75, further analysis suggested that this might be due to the small number of samples in the arterial phase (1–2 cases), resulting in a higher chance agreement and thus a lower overall Kappa value. Therefore, this feature was retained (Table S4 and S5). In total, 27 features were included in the study. Given that only 42 patients with pMMR and pT4 staging required NAC followed by surgery and postoperative chemotherapy, the 10 EPV rule was applied, allowing for a maximum of four features to be retained in each task layer. Additionally, there was some overlap in the features retained across different task layers. The total feature set retained by five feature screening algorithm is shown in Table S6.

Model performance evaluation
The total feature set obtained from five feature screening methods was applied to three tasks using 14 machine learning algorithms, evaluated through five-fold cross-validation. The results indicate that the total feature sets derived from Lasso, mRMR, and XGBoost-RFE methods achieved average AUC values greater than 0.8 across all validation datasets (Fig. 4a). Since the MMR task is the initial step in global decision stratification, we utilized the validation results of this task as the primary evaluation metric for further feature screening methods selection. The combinations of Lasso + Logistic Regression, mRMR + SVM, and XGBoost-RFE + SVM achieved the highest AUCs in the validation set for the MMR task, with values of 0.79, 0.77, and 0.83, respectively. When evaluated on the test set, these models yielded AUCs of 0.69, 0.80, and 0.66, respectively. Consequently, the final total feature set was based on the mRMR feature screening method, and the final machine learning algorithm for the MMR task was SVM.

The total feature set based on the mRMR method was then used for testing in the pT task and risk factors task. The results revealed that the Bernoulli NB algorithm achieved the highest AUC of 0.82 in the pT task test set, while the Decision Tree algorithm had the highest AUC of 0.85 in the risk factors task test set. Detailed results for the final models of all three tasks are presented in Table 1. Figure 4b presents the ROC curves of test set for the models of three tasks. In addition, the models developed for the three tasks are accessible through three dedicated websites (Supplementary Material Ⅱ). This allows for real-world testing and application, enabling users to evaluate the models with additional datasets.

Furthermore, the direct multiclass model constructed using the total feature set derived from mRMR screening, in conjunction with the Bernoulli NB algorithm, exhibited optimal performance, achieving an average AUC of 0.77 across four classifications in the test set (Figure S1 in Supplementary Material Ⅱ). Both model types demonstrated high accuracy in predicting the type of surgery combined with postoperative chemotherapy. However, when evaluated using confusion matrices (Fig. 4c and d) and weighted MCC, the multitask hierarchical model developed in this study yielded result of 0.42 for global decision stratification, which was notably higher than the weighted MCC of 0.32 for the direct multiclass model.

Discussion
Accurate stratification of NMCC patients is crucial for tailoring individualized treatment strategies. While previous studies have used traditional statistical methods, radiomics, or deep learning to predict MMR status [15, 20], pT stage [21, 22], N stage, or other pathological risk factors [14, 23], this study integrates conventional CECT signs and clinical features to develop a novel multitask machine learning model. This model predicts MMR status, pT stage, and high-risk status, facilitating preoperative global decision stratification for NMCC patients. Multitask predictions can capture shared features of high-risk factors, reducing the complexity of stepwise stratification. Our results show that the feature set of 12 variables performs well in predicting MMR status, pT stage, and high-risk status, contributing to effective risk stratification in NMCC patients. It should be noted that this framework does not constitute the exclusive approach to clinical decision-making. In clinical practice, surgical decision-making is influenced by multiple factors, including tumor staging, lymph node metastasis, presence of intestinal obstruction or perforation, among others. However, our risk stratification framework serves to facilitate preoperative decision-making, optimizing clinical outcomes for patients.
In this study, we focused on analyzing CECT images (both arterial and venous phases) to evaluate tumor and lymph node characteristics, as these phases provide critical hemodynamic information about tumor vascularity, perfusion, and tissue heterogeneity. While unenhanced CT images can provide baseline information (e.g., tumor necrosis or calcification), dynamic CECT offers more comprehensive insights beyond baseline data. The feature set of the final constructed model incorporated nine CECT signs and three clinical features. Notably, among the nine CECT features, seven were directly associated with the primary tumor characteristics of colon cancer (Tumor-Max-VP, Tumor length, Location, Necrosis, Tumor-Mean-VP, Enhancement characteristics, Growth pattern). This finding underscores the critical importance of evaluating baseline CT features of the primary tumor in radiological diagnosis. Our model revealed that elevated that both Tumor-Max-VP and Tumor-Mean-VP values — preferentially associated with early T-stage tumors (pT1–3), contradicting historical studies linking hyperenhancement to poor prognosis [24]. This may stem from dichotomous biological meanings of angiogenesis across tumor evolution stages [25]. The hyperenhancement observed in early-stage tumors may be attributed to physiological angiogenesis, which supplies oxygen and nutrients to support tumor growth and sustains tumor differentiation, manifesting as localized growth. In contrast, hyperenhancement in advanced tumors arises from pathological angiogenesis driven by hypoxia-induced overexpression of vascular endothelial growth factor (VEGF)/VEGF receptor 2 (VEGFR2). This leads to disorganized vascular architecture and leakage, which are closely associated with tumor invasiveness.
Research findings demonstrate that larger tumor size, tumor necrosis, and heterogeneous enhancement are associated with dMMR status or advanced tumor stage (pT4). Notably, these three features often indicate higher malignancy and poorer prognosis. However, previous studies have shown that dMMR colorectal cancer patients have better prognoses than pMMR patients [26]. In line with this, our study found that pMMR patients had higher risks for N stage, differentiation grade, and perineural invasion compared to dMMR patients. However, we also observed that dMMR patients had a higher proportion of primary tumor necrosis and heterogeneous enhancement, possibly due to increased immune infiltration caused by the mismatch repair deficiency, which leads to an influx of immune cells into the tumor microenvironment, thereby inhibiting tumor growth and spread [27, 28]. Additionally, the release of cytotoxic factors or inflammatory mediators from immune cells in the tumor microenvironment may cause local necrosis, which manifests as heterogeneous enhancement on CECT. Similarly, dMMR patients showed greater lymph node burden, likely due to immune-driven reactive hyperplasia rather than metastasis [29]. In contrast, the increased lymph node burden in pMMR patients with pT4 stage or high-risk pathological factors is more likely a sign of true tumor metastasis due to the higher local invasiveness and metastatic potential of these tumors. We believe that future research will increasingly focus on differentiating reactive from metastatic lymph nodes in dMMR and pMMR tumors, which could enhance clinical utility.
This study employed various feature screening algorithms to construct a total feature set, enabling multitask predictions based on a unified set of features for global decision stratification. All five feature screening methods retained the CECT signs of necrosis, obstruction, and perforation, indicating their significant clinical relevance in assessing the prognosis and decision stratification of NMCC patients, which is highly consistent with previous research [30–32]. Notably, the mRMR algorithm preserved 12 conventional features. With the exception of the enhancement characteristics and growth pattern in the risk factor task, which differed from the univariate analysis results, all features in the MMR task and pT stage task, as well as the obstruction and CEA features in the risk factor task, aligned with the statistical analysis findings, further confirming their close association with outcomes. The discrepancies between enhancement characteristics and growth pattern in statistical versus machine learning feature screening may stem from the latter’s approach, which optimizes global predictive performance by considering the importance and contribution of each feature within the model, as well as the interactions between features. The goal of these methods is to enhance model performance rather than merely identifying features that are statistically significant with the dependent variable. In contrast, traditional statistical methods often overlook feature interactions or redundancies, meaning that certain statistically significant features may be highly correlated with other features or contribute little to the overall prediction, leading to their exclusion in feature selection.
The multitask hierarchical model developed in this study demonstrates excellent predictive performance across three tasks: MMR, pT stage, and risk factors. For MMR prediction, the SVM excels in handling high-dimensional data and nonlinear classification. In contrast, for pT staging and risk factors, the interpretability of the Bernoulli NB and Decision Tree algorithms enhances the understandability of the results. This multitask learning strategy not only improves the predictive performance of each task but also effectively reduces the risk of overfitting and enhances the model’s generalization capabilities. By considering multiple related tasks simultaneously, the model captures the latent relationships between tasks, thereby improving predictive accuracy. When conducting final global decision stratification, our multitask hierarchical model outperformed direct multi-class model, highlighting its strengths in integrating multidimensional information and enhancing predictive precision. However, the overall accuracy of the model still requires improvement, which may be attributed to “error propagation” within the hierarchical framework [33]. Future research will explore strategies such as deep ensemble learning, error correction mechanisms, and uncertainty assessment to optimize performance further. Importantly, the study constructs models based on common CT morphological features, which possesses interpretability and reproducibility and is more suitable for clinical application. The model provides independent predictive results based on comprehensible conventional clinical features and CECT signs, enabling clinicians to gain clearer insights into tumor characteristics and overall patient risk. This approach aligns more closely with real-world clinical reasoning and decision-making pathways, thereby enhancing the model’s interpretability.
This study has several limitations. First, being retrospective and excluding lesions that could not be identified, they carry inherent risks of selection bias. Second, the sample size of T4 patients is relatively small, and the sample imbalance may affect the accuracy of the findings. Although we attempted to use SMOTE oversampling to balance the data, it led to overfitting, so this approach was abandoned. Future studies with larger, multicenter cohorts are needed to validate these findings and improve model performance. Third, differences in patient enrollment time windows between the two institutions may introduce potential bias. However, given the inherent stability and standardization of the core radiologic and clinical features, we believe the temporal discrepancy between centers does not meaningfully compromise data integration, analytical validity, or the reliability of our conclusions. Fourth, this study did not utilize non-contrast CT features. While dynamic CECT provides hemodynamic information, non-contrast CT features may offer complementary value in assessing tumor heterogeneity. Future studies could integrate non-contrast CT, multiphase enhancement data, and clinical parameters to develop more holistic predictive models. Besides, this study lacks external validation. Although the robustness of the model was confirmed through 5-fold cross-validation, future multicenter studies employing standardized CT protocols are warranted to evaluate its generalizability. Besides, this study did not incorporate non-contrast CT features, but employed dynamic CECT protocols encompassing both arterial and portal venous phases, thereby offering more comprehensive insights beyond baseline data. Lastly, this study did not include a survival analysis, which limits our understanding of the long-term outcomes and prognostic value of the identified features. Future research incorporating survival analysis will be essential to provide a more comprehensive assessment of the clinical utility of the proposed models.

Conclusion
In conclusion, we presented a multitask machine learning approach to achieve global decision stratification in NMCC patients based on MMR status, pT stage, and high-risk pathological factors, utilizing both CECT imaging features and clinical parameters. Our results demonstrated that the multitask hierarchical model can effectively predict key clinical outcomes, providing a non-invasive preoperative tool for NMCC patients. This approach offers new insights and methodologies for personalized medicine, and it holds promise for significant impact in future clinical practice.

Supplementary Information
Below is the link to the electronic supplementary material.

Supplementary Material 1